DeepSeek-Coder-Instruct 6.7B
DeepSeek · 2023-11-02 · 6.7B parameters
DeepSeek-Coder-Instruct 6.7B is the instruction-tuned version of DeepSeek-Coder-Base 6.7B from the original DeepSeek-Coder release of November 2023, fine-tuned on about 2B tokens of instruction data. It keeps the base model's 6,740,512,768-parameter dense architecture and 16K context.
Benchmark scores
| Benchmark | Score |
|---|---|
| ARC-Challenge | 37.4 |
| BBH (BIG-Bench Hard) | 46.9 |
| Capability | 88.7 |
| HellaSwag (Unspecified Scoring Protocol) | 55.0 |
| HumanEval | 78.6 |
| HumanEval-Mul | 66.1 |
| MBPP | 65.4 |
| MMLU | 37.2 |
| WinoGrande | 57.6 |